You are on page 1of 23

Journal of Fmancml Economics 8 (1980) 31-53 Q North-Holland Pubhshmg Company


Market-Making with Inventory*

Tel-Avlv Umversrty, Tel-Avru, Israel
Columbia Umuersrty, New York, NY 10027, USA

Unrversrty of Rochester, Rochester, NY 14627, USA

Received July 1979, final version received January 1980

This study considers the problem of a pruze-settmg monopohstlc market-maker m a dealershlp

market where the stochastic demand and supply are depIcted by prlcedependent Poisson
processes [followmg Garman (1976)] The crux of the analysis IS the dependence of the b&ask
prices on the market-makers stock Inventory posltlon We derive the optlmal pohcy and Its
characterlstlcs and compare It to Garmans The results are shown to be consistent with some
coqectures and observed phenomena, hke the existence of a preferred mventory posItIon and
the downward monotomclty of the bid-ask prices For hnear demand and supply functions we
derlvethe behavior of the bid-ask spread and show that the transactlon-to-transactlon price
behavior 1s mtertemporally dependent However, we prove that It IS lmposslble to make a profit
on this price dependence by tradmg agamst the market-maker Thus, m this sltuatlon, serially
dependent price-changes are consistent with the market etliclency hypothesis

1. Introduction

The microstructure of non-Walraslan dealership markets IS a SubJect of

growing interest The mam issue m question IS the impact of the actlvltles of
a market-maker, acting on his own behalf (SubJect to institutional and ethical
constraints) on the operational characteristics of the market
In a pioneering study, Garman (1976) presented a rigorous stochastic
model of the dealershlp market This dealership market 1s entirely dominated
by a centralized market-maker, who possesses a monopoly on all trading
Being a price-setter, the market-maker quotes bid and ask prices that affect
the stochastic mechanism which generates market sell and buy orders,
respectively Garman introduced a stochastic analogue of the classic supply

*The authors are grateful to Avraham BeJa, Michael C Jensen and the referees, Robert
Wilson and Peter Kubat, for helpful comments and suggesttons
32 Y Amrhud and H Mendelson, Market-making with inventory

and demand functions which makes the operation of this mechanism

tractable He suggested that the collective actlvlty of the market agents can
be characterized as a stochastic flow of market sell and buy orders whose
mean rate per-unit-time 1s price-dependent This gives rise to a market
supply (demand) curve which depicts the expected mstantaneous arrival rates
of mcommg sell (buy) orders as a function of the quoted bid (ask) price
The possible temporal discrepancy between market buy and sell orders,
and the obhgatlon to mamtam contmuous trading, induce the market-maker
to carry stock inventories, either positive (long posltlon) or negative (short
position) Garman studied the lmphcatlons of some inventory-independent
strategies, which are based on the selection of a fixed pair of bid-ask prices,
and showed how they lead either to a sure failure or to a possible failure (see
also m the next section) He suggested that the specialists must pursue a
pohcy of relating their prices to their inventories m order to avoid failure
[Garman (1976, p 267)] This Inventory-dependent pohcy is, m fact, the
main issue of our paper
In this study we derive the optimal prlcmg pohcy of the market-maker m a
Garman-like dealership market, SubJect to constramts on his short and long
stock inventory posltlons The crux of the analysis 1s the dependence of the
quoted bid and ask prices on the market-makers stock We derive the
optimal pohcy and show that its characteristics are consistent with some
conjectures and observed phenomena It 1s proved that the prices are
monotone decreasing functions of the stock at hand, and that the resulting
spread 1s always positive It 1s shown that the optimal pohcy implies the
existence of a preferred inventory posltlon, as was suggested by Smldt
(1971), Barnea and Logue (1975) and Stoll (1978a) We also obtain some
noteworthy relations between Garmans model and ours concernmg the
ObJective function values and pohcy variables
Focusing on the case of linear demand and supply, we derive the explicit
behavior of the bid-ask spread and the expected tradmg volume as functions
of the inventory posltlon Most importantly, we prove that the optimal
pricing pohcy 1s consistent with the efficient market hypothesis m the sense
that it 1s lmposslble to make a profit by speculating m the market (except, of
course, the market-maker, who enjoys a monopohstlc position) Thus, a
transaction-to-transaction price behavior which lacks mtertemporal
independence may well be consistent with the market etliclency hypothesis
Recently, there has been a growmg interest m the posslblhty of
computerlzmg part of the market-makers functions m the securities markets
[e g, BeJa and Hakansson (1979), this idea goes back to Fama (1970)] This
Barnea and Logue (1975) also argued for an Inventory-dependent prwng pohcy Stall
(1978a) dewed the effect of Inventory holdmg costs on the market-makers quoted prices BeJa
and Hakansson (1979) consider the lmphcatlons of some Inventory-dependent tradmg rules for
Y Amlhud and H Mendelson, Market-makmg wtth mventory 33

requires the specllicatlon of a transaction-by-transaction pricing pohcy for

the market-maker We hope that our study 1s a step towards the feasible
application of this idea
It has been suggested by Bagehot (1971) that the market-maker IS faced
with basically two kinds of traders the hquldlty-motivated transactors, who
do not possess any mformatlon advantages, and insiders which are
transactors with superior mformatlon He suggested that the market-maker
gains from the former and loses to the latter, and the tradeoff between the
two determines his spread It should be noted that our dealership market
(and Garmans) 1s intended to describe the hqmdlty-motivated transactions
It IS worth mentlonmg some other avenues of research pursued in the
study of market-maker impact on the securities markets The role of the
market-maker as a price-setter and stabilizer m the stock market was
presented by Baumol (1965, ch 3) who studied the effect of his monopohstlc
posltlon on market prices and resource allocation In a recent overview of
market mechanisms, BeJa and Hakansson (1977) discussed some alternative
means to facilitate trading m the securltles markets The impact of trading
mechanisms on various characterlstlcs of price behavior m dealership
markets was investigated m a series of studies by Cohen, Maler, Schwartz,
Whitcomb, and Ness and Okuda In particular, they examined the effect of
market thinness on the moments of stock returns m securltles markets which
operate with or without central market-makers [Cohen, Ness, Okuda,
Schwartz and Whltcomb (1976), Cohen, Maler, Ness, Okuda, Schwartz and
Whltcomb (1977)], furnished a theoretical framework for these phenomena
[Cohen, Maler, Schwartz and Whltcomb (1978a)], and suggested pohcy
lmphcatlons [Cohen, Maler, Schwartz and Whltcomb (1977a)] Cohen,
Maler, Schwartz and Whltcomb (1977b) explained the existence of serial
correlation m the securltles markets, even when the generated quotations
have zero own- and cross-serial correlation, by the existence of bid-ask
spread and non-simultaneous transactions in various stocks BeJa and
Goldman (1977) showed how the way in which expectations are formed may
affect the rate of price convergence to the equlhbrlum price BeJa and
Goldman (1978) also showed how the impact of a speclahst may lead to a
serial correlation m security returns although the underlying process 1s a
random walk
In a simulation study, Bela and Hakansson (1979) assessed the
ramlficatlons of various trading rules adopted by a (programmed) speclahst
whose role 1s to smooth the discrete demand function by buying or selling a
sufficient number of shares to clear the market
Another group of studies [e g , Demsetz (1968), West and Tmic (1971),
Tmlc (1972), Benston and Hagerman (1974), Barnea and Logue (1975),
Logue (1975), and Stoll (1978b)] focus on the determinants of the bid-ask
spread charged by the market-maker They suggest that the bid-ask spread 1s
34 Y Amthud and H Mendelson, Iclarke&makmg wath mventory

a function of the cost of provldmg immediacy services, the risks to the

market-maker inherent m the traded stock (m particular the specific risk due
to insiders trading) the degree of competltlon, the depth of the market, the
price per share and the volume of trading
The notion of a dynamic price-inventory adlustment policy was discussed
by Smldt (1971, 1979), Barnea (1974), Barnea and Logue (1975) and Stoll
(1976) They suggested that the market-maker has a preferred inventory
position and when his realized inventory deviates from It - he will adjust
the level, and possibly the spread, of bid-ask prices to restore that position
Stoll (1976) presented and tested a model of dealer inventory response to
past, current and future price changes and found that speclahsts tend to buy
stocks on price declines and sell stocks on price increases In a later model,
Stoll (1978a) considered an expected utility maxlmlzmg dealer whose quoted
prices are a function of the cost of taking a posltlon which deviates from his
desired position, and derived the lmphcatlon of his inventory pohcy on the
bid-ask spread and on the structure of the dealership market
In what follows, we present the model m section 2, derive the optimal
pohcy and its characteristics m section 3, and solve for the case of linear
demand and supply m section 4 In section 5, we discuss some additional
aspects and possible extensions

2. The model
The market considered m this study is slmllar to the dealership market
introduced by Garman (1976) The mam features of this market concern the
monopohstlc position of the market-maker and the nature of the aggregate
supply and demand functions Followmg Garman (1976, p 263), the
underlying assumptions on the market are

(A) All exchanges are made through a single central market-maker, who
possesses a monopoly on all trading No direct exchanges between
buyers and sellers are permitted
(B) The market-maker 1s a price-setter He sets an ask price, P,, at which
he will fill a buy order for one unit, and a bid price, P,, for a one-unit
sell order
(C) Arrivals of buy and sell orders to the market are characterized by two
independent Poisson processes, with arrival rates D(P,) and S(P,),
respectively The stationary price-dependent rate functions D( ) and
S( ) represent the market demand and supply, with D( ) < 0, S( ) > 0
(D) The ObJectWe of the market-maker is to maximize his expected average
profit per unit-time Profit 1s defined as net cash inflow
Y Amlhud and H Mendelson, Market-makmg wrth mentory 35

These assumptions were used by Garman to explore the behavior of the

market-maker He derived the probablhtles of failure and the necessary
condltlons to avoid a sure failure Then he modified the assumptions to
analyze two cases In the first, the market-maker sets bid and ask prices to
maximize his expected profit per umt time subject to the constraint of no
inventory drift In the second case, Garman assumed zero price-spread and
analyzed the nature of failure and the duration of the process
Our model IS a natural extension of Garmans As he suggested, we allow
the prices set by the market-maker to depend on his stock inventory
posltlon, which leads to a dynamic pricing pohcy of the market-maker For
that purpose we focus on the stochastic process which describes the
development of inventory This process results from the arrival of market
buy and sell orders whose rates are governed by the pricing declslons of the
market-maker The underlying process of market-orders generatlon m our
model 1s identical to that assumed by Garman The arrival processes are
Poisson processes whose rates are price-dependent Yet, since our concern 1s
with the development of inventory, we shall rewrite Garmans assumption
(C) m a form suitable to our purpose
It 1s well known that the Poisson process IS characterized by independent
exponentially distributed mterarrlval times Thus, a given pan of prices, P,
and P,, generates two competmg exponential random variables z, with
mean l/D(P,), and z,, with mean l/S(P,) [see e g Howard (1971, pp 793-
797)] The next arriving order will occur at time mm{z, z,>, being a buy
order d z, c z,,, or a sell order if z, > TV It follows that assumption (C) may
be written m the followmg equivalent form

(C) For a given pair of pnces, P, and P,, the next mcommg order will be a
buy order with probablhty D(P,)/(D(P,)+S(P,)), or a sell order with
probability S(P,)/(D(P,)+S(P,)) The time until the next arriving
order has an exponential dlstrlbutlon with mean l/(D(P,) + S(P,))

Formulation (C) lmphes that the process of mventory development (due

to the discrepancy between supply and demand) IS m fact a birth and death
process [see, e g, Howard (1971, pp 797-814)], whose parameters are
controlled by the market-maker Thus we obtam a semi-Markov decision
process [Howard (1971, ch 15)] where the state variable 1s the stock at
hqnd, and the decision made for a given mventory level 1 is a pair of prices,
P,, and P,, which determine the respective demand and supply rates
The stock at hand 1s assumed to be bounded from above by some
constant L and from below by -K, where K and L are integers and -K c L
This assumption reflects the hmltatlons which are usually imposed on the
market-makers ablhty ot take long and short positions These hmltatlons
result from capital requirements or from admmlstratlve rules Note that the
36 Y Arnlhud and H Uendelson. Market-makmg with mventory

formulation of the model m terms of this finite state-space resolves the

problem of possible rum Formally, we assume
(E) The permissible stock inventory levels are {-K, -K + 1, -K +2, . ,
L-2, L- 1, L}
For convenience of exposltlon, we re-number the states as (0,1,2, ., M
- 1, M}, where M= L+K We assume M 2 3. We also adopt the
conventional termmology o&rth and death processes, and let rl, denote the
birth rate m state k, and R the correspondmg death rate. We also define cl0
=A, =0 Since A, = S(P,) 1s a monotone mcreasmg function of PbL, there IS
a one-to-one correspondence between 1, and Pbk Simlarly, pk IS a monotone
decreasing function of Pak, with a one-to-one correspondence Thus, the
transition rates I, and pk (rather than the correspondmg prices) ~111be used
as the decision variables m state k
The market demand and supply ft&l&s give rise to the market-makers
revenue and cost functions, respectively,

W)=P P.OI)=P 0-01),

C(i)=J P,(rz)=I2 s-(n)

Rb) represents the expected sales revenue per unit time corresponding to
demand rate p (which IS, m turn, a function of the pre-determined ask price
Pa) Analogously, C(n) 1s the expected cash outlay per unit time, which
represents the cost of inventory replenishment
We make the followmg regularity assumptions on R( ) and C( )

(F) The market-makers revenue and cost functions, R( ) and C( ),

respectively, are twice contmuously differentiable with
(1) R( ) is strictly concave, I e , R(p)<O,
(ii) C( ) is strictly convex, 1e , C(Iz)>O,
(III) R(O)>C(O), R(oo)cC(co)

Finally, we assume
(G) There are no transaction costs to the market-maker
Note that the existence of a transactlon cost 5 per trade [see Garman (1976,
p. 266)] paid, e g , by the market-maker, will simply shift the supply and
demand functions by a constant, retammg the convexity and concavlty
properties of C( ) and R( ) Thus, there IS no loss of generality m assummg
<=O (Obviously, other costs which are constant per unit time do not affect
the optimal pohcy )
Y Amthud and H Mendelson, Market-makmg wrth Inventory 37

The problem is to find the optimal pohcy of the market-maker under

assumptions (A)-(G) Followmg assumption (D), the ObJectWe function IS
[Howard (1971, p 868)]



A=(&, ,A,-,) and ~=(PI, ,PM)

& is the earning rate for a transition from state k, I e , the expected cash flow
dlvlded by the mean sojourn time In terms of our model, the expected cash
flow per transition from state k 1s

1k R@k)-C(Ak)
--&- pabk)- - Pb(Ak)=
Ik+pk Ik+pk Ak+pk

and the expected SoJourn time in State k is (& + pk.- , hence


+I, is the limiting probablhty of finding the process in state k It is well

known that the stationary probablhtles 4k (k =O, 1, , M) satisfy the relations

Ak+k=~k+l~k+l, k=O,l, ,M-I, (3)

that is, the expected flow from state k to state k+ 1 equals the expected flow
m the opposite direction It follows that




be the mean rates of mcommg sell and buy orders Then (3) lmphes

X=ji, (6)

that is, the expected flows In both directions are equal Relations (3) and (4)
are valid only when A,>0 for k=O,l, ,M-1 and pk>O for k=l,2, ,M
38 Y Anuhud and H Mendelson, Market-makmg with menlory

It will be proved m the sequel that under the optimal pohcy, the birth and
death rates are Indeed positive

3. Optimal market-maker behavior

In this section we first derive the optlmahty condltlons and then study
then lmphcatlons on the behavior of the market-maker By a stralght-
forward dlfferentlatlon of the obJectlve function (1) we obtain the followmg
necessary condltlons for optlmahiy


pk f ~,CR(~,)--((I,)l--k~kR(~k)=g~,C1)
J=k ]=k

By subtracting the (k + 1)st equatron of (8) from the kth equation of (7)
and using (3) we obtain

R(&+l)=C(&), k=O,l, ,M-1, (9)

which remmds of the ordinary optlmahty condltlon of a monopoly, except

that here it relates to each pair of nelghbormg states Note that since

PaOLk+l)>R(~k+~)=C(~k)>Pb(~k), (10)

a purchase of one unit at state k and Its sale at state k+ 1 always yields a
profit z It follows that a loop of transitions starting from any state k,
traversing other states and returning to state k yields a posltlve profit with
probability one Thus, when the market-makers mltlal resources (cash plus
available credit) exceed XL-,-,P,,, the probability of cash failure [Garman
(1976, p 263)] is zero, even m the worst possible case Since the probability
of default by the market-maker 1s zero, mltlal credit (If needed) should be
Subtraction of (7) from (8) yields

c(n,)-n,c(n,)+g~,~)=O, (IlO)
c(n,)-n,c(nk)+g~,~)=ROI,)-~kR(CIk), (Ilk)

k=l,2, ,M-1,

ZA sale of a umt at state k+ 1 can always be attributed to a purchase at state k (except for
the mtlal stock)
Y Amrhud and H Mendelson, Market-makmg wrth mventory 39

g@4.4=wMhdwhf) (119
Eqs (11) give the basic relation between A, and Pi, and thus the relation
between the bid and ask prices for each inventory posltlon k It follows that
the optimal (A,,p& are aligned along the curve defined by

[note that @,,O) and (0,~~) are at the intersection of the curve with the
positive semi-axes] Lemma 3 1 proves that the curve (12) depicts a
downward sloping function on the positive quadrant

Lemma 3 1 Along (12), for A,p>O, dp/dAcO

Proof The proof follows smce for A, p > 0,


<o QED (134

The followmg theorem establishes that the optimal bid and ask prices are
monotone decreasing functions of the stock at hand [This result IS consistent
with the dynamic pricing pohcy described by Smldt (1971, 1979), Barnea
(1974) and Barnea and Logue (1975) ]

Theorem 3 2 Let P,, and Pak, respectively, be the optimal bid and ask prices
at state k Then P,,>P,, >P,,> >P,,,_, and P,, >P,,> >PaM
Equtvalently, 1, > A, > I, > >A,-, andpl<p2<p3< <PM

Proof The proof IS given in terms of 4 and p_ The equivalent formulation

m terms of price behavior follows from the moiotomclty of the demand and
supply functions
We first prove that i, >A, By subtracting eq (1 lo) from (11) and noting
that R(c(~ )/pl > R(pl ). we obtam

The inequality now follows from (13b)

40 Y Amthud and H Mendelson, Market-makmg wzth mwntory

Next, &,>A, tmphes C(A,)>C(A,), hence by (9), RI@,)> R(p2) It

follows that p, <pz, and then we obtam by Lemma 3 1 that A, CA, The
proof IS completed by mductlon Q E.D

The following corollary shows that the bid-ask spread IS always posltlve,
as might be expected

Corollary 3 3 Pak- P,, > 0

Proof By (10) and Theorem 3 2 we obtam

p.k=p*o1k)p,o1k+l)~p,(~k)=p,k QED
We now turn to compare our results to a case treated by Garman (1976,
pp 265-266) Consider a market-maker who wishes to prevent a drift m his
expected inventory Thus he sets prices so as to equate the rates of mcommg
buy and sell orders, 1e , p =A, regardless of his inventory posltlon This
market-maker acts like an ordinary monopohst who equates marginal
revenue to marginal cost
It might be argued that the profits of this monopolist, who restricts himself
to p=I, are lower than those of our market-maker who enjoys a greater
flexlbdlty m setting prices Yet, our market-maker has constraints on the
long and short posltlons which he can take, whereas Garmans monopolrst
has no such constraints In fact, the followmg theorem proves that the profit
of Garmans monopolist 1s an upper bound on g&p)

Theorem 3 4 Let I* be the (unique) solutton of R(r*)= C(r*) Then

g@,pli)<R(r*)-CO-*) (14)

Proof The ObJective function of our market-maker 1s

g&p)= 5 4rCRW-C(&)1

The function R(p)- C(A) is strictly concave on the @,I)-plane, hence, by

Jensens Inequality,

where p and X are given by (5) Furthermore, it follows from (6) that ji=X 1s
a subset of the constraints m our problem Now, Garmans problem can be
Y Amahud and H Mendelson, Market-makmg wtth anventory 41

written as

st j.l=A,

whose solution IS Iz=p = I* It follows that for all 2,~

g~,~)<R(jl)-c(X)~h(r*,r*)=R(r*)-C(r*) QED

We now employ Theorem 3 4 to prove that It IS not optimal to have a

vamshmg transitian rate That K., the profit-maxlmmng market-maker will
never choose to refrain from makmg buy or sell transactions (except for the
case where he reaches his hmltmg posltlons) This also means that relaxing
his constraints by expanding the allowed short or long posltlons strictly
Increases the market-makers profits

Theorem 3 5 The optimal policy (A,& satisfies I, >O for k=O, 1, , M - 1,

and pn>Ofor k=l,2, ,M

Proof It 1s sufficient to prove that ~1~=0 IS not optimal 3 For that purpose
we apply Howards (1971, pp 983-1005) pohcy improvement procedure to
show that a pohcy with p1 >O 1s an improvement over the pohcy with pL1=0
Let the mltlal pohcy (2,~) be such that pL1=0 and A, = I* (note that smce
state 0 IS transient, the value of I, does not affect g), with relative state
values Us Consider an alternatlve pohcy @,g) where A=A, & = p, for all
J# 1, but & = r*/2>0 Let r1 be the value of the test quantity [Howard
(1971, p 986)] for evaluating the orlgmal pohcy at state 1, and P1 the
correspondmg test quantity for the alternative pohcy Then,

r,=-c(n,)+(n,+o)[l Q-U,],

[ :;*,2v1+A

, 1

r;-r,=R(r*/2)-r*/2 (ul-u())

More generally, the proof ImplIes that addmg one state to a cham strictly Increases the value
of g (m that chain) Hence, the cham resultmg from the optlmal solution must contam all states
42 Y Amrhud and H Mendelson, Market-makmg wzth mventory

Now, (IJ~-u,,) IS found by performing the pohcy evaluation procedure for the
initial policy at state 0,

uo+ co = - wo )l~, + 01,

that IS,

l-00 =CdL&+ C(r*)llr*

By Theorem 3 4, g@,,r)<R(r*)-C(r*), hence u1 -oo<R(r*)/r*, and

r;-f,>R(r*/2)-R(r*)/2>0 QED

We now proceed to study the characterlstlcs of the stationary dlstrlbutlon

{&},=o under the optimal pohcy It follows from Theorem 3 5 that & > 0 for
all k=O,l, , M, so there 1s a posltlve probablhty of finding the market-
maker m any of the allowed inventory positions In addltlon, other
properties of {&},=o are of interest What IS the shape of this dlstrlbutlon7
Is there a preferred inventory posltlon [Smldt (1971, 1979), Barnea (1974),
Barnea and Logue (1975), Latank et al (1975), Stoll (1976)] If there IS, what
can be said about the preferred rates and pnces.7
Smce ~k+l/4k=~J~k+19 the properties of {&},=, may be obtained by
consldermg the transition rates 2, and /A~+~ The followmg lemma relates Ak
and pk+l to Garmans no-dnft rate r*

Lemma 36 For all k=O,l, ,M-1,

with equality if and only tf 1, = r* =pk+ 1

Proof (I) If pk+l c(=)r*, then

C'(&)= R(pk+ 1) > (= )R(r*) = C(r*),

hence I, > ( = )r*, respectively

(II) If pk+ I > r*, then


hence I, -Cr* QED

Lemma 3 6 states that r* IS always located between i, and pk+ 1 This fact
IS Illustrated m fig 1
Y Amrhud and H Mendelson, Market-making wth mventory 43

Theorem 3 7 The dlstrrbutlon {c#+},=, 1s urnmodal The mode J 1s such


I,zr* for k<J, l.,<r* for kzJ, (15)

pkSr* for kSJ, pk>r* for k>J (16)

Proof First, 1, > r* smce otherwise we would have A, ~1,s r* 6p1 < pk+ 1
for all k= 1,2, , M - 1 This would have lmphed x<p, m contradlctlon to
6) A slmdar argument leads to py > r* It follows from Lemma 3 6 that

h+l Xk
Expected number of buy (~1 and sell (A) orders
per unit time

Fig 1 The market-makers revenue as a function of the arrival rate of buy orders, R(P), and his
replemshment cost as a function of the arrival rate of sell orders, C(A) r* IS the arrival rate
chosen by Garmans monopohst, maxlmlzmg R(r)- C(r) 2, and pr+ 1 are related by (9) and
thus the Interval [pt+ 1,AJ embraces r*

Next, &/& _ 1= & _ Jpk, which - by Theorem 32 - IS a strictly

decreasmg function of k Let


Inequalities that J 1s well defined and O<J< M For k$J,

(17) imply
4k2$Jk-19 whereas k > J, & < c$_ 1 It follows that 4. <
for <4,-2-C
4J-1 S4,P and $J>4J+1>4J+z> >& Now, (15) and (16) follow
from Lemma 3 6 and from Theorem 3 2 QED

This also Includes the case of two consecutive modes, J and J - 1

44 Y Amrhud and H Mendelson, Marketqakmg wrth rnventory

Theorem 3 7 suggests that the market-maker adopts a prlcmg pohcy whmh

produces a preferred Inventory position J, located away from the hmltmg
positions 0 and M The preference for posmon J 1s reflected m both the
ummodahty of the dlstrlbutlon and m the fact that the decline rate of & as k
withdraws from J, IS faster than a geometric decline rate (since 4 J&-l IS
decreasing) Furthermore, when the market-maker finds himself m a posmon
different from J, he will quote prices which will tend to bring him back to
that posmon. That IS, the probablhty that the next transition will be m the
due&on of J will exceed the probablhty of moving towards the extreme
positions This aversion from the extremes follows since being there forces
the market-maker to make transactions at unfavorable condmons
Furthermore, at the preferred inventory posmon J, L and p are
approximately equal An exact equality between I and /1 IS obtained at
(1,&, where the curve (12) intersects with the 45-lme (see fig 2) The actual
(AJ,pJ) IS located m the neighborhood of (A,$), m the followmg sense
Let a = mm (A,, pJ}, and A = max {A,,pJ} Then,


(a,A)c(~,+,,C1,+1)C(~,+,,~,+2)C = (O,Phf),
where a51=pS A
Thus, all the intervals between 1, and pLkstraddle the interval between 1,
and p,, which m turn straddles A=pl This gives (A,pi) the mterpretatlon of
being the approximate likely rates of the market-maker If some I, happens
to equal 2, then k = J IS the preferred inventory posmon and A, = I =P = pJ
are the exact modal rates
We finally relate the likely rates (A,$) to (r*, I*), the rates of Garmans

Theorem 38 d=p<r*

Proof Let h(x)=[R(x)-xR(x)] -[C(x)-xc(x)] Now, h(x)>0 for x>O,

h(r*)=R(r*)-C(r*), and by (12), h(l)=g It follows from Theorem 3 4 that
h(l)<h(r*), hence 1=p<r* QED

The above relation, together with (15) and (16), are summarized m fig 2
Note that both (n,,pL,) and (A,$) he on the segment slsl Furthermore, the
only (&,p,J contamed m this segment are the modal rates More specifically,
when I,_,/~J>l, we have ;1,_,>r*>p,>p,_, and 2J+l<<J<r*<pJ+1
Y Amrhud and H Uendelson, Market-makmg with mventory 45

X, Expected number of sell orders per unct time

Fig 2 The curve [R(p)-pR(p)] - [C(n)-X(i)] =g, along which are ahgned the optunal
arrival rates of the pubhc buy and sell orders br and I,, respectively, with the subscrlpt
denotmg the related Inventory posItIon) The hkely rates sattsfy 1=~ The actually preferred
rates (A,,p,) are both less than or equal to Garmans rates (r*.r*), thus they are contamed m
the segment s,s2

Hence, both modal rates are smaller than the rates of Garmans monopohst,
I e , (A,,pJ) IS mslde s1s2, and for k#J, (A,,p,) hes outslde slsl In the case
where 1,-,/p,= 1, (Ar-l,~,_,) and (A,,p,) comclde with the endpomts s1
and s2, respectively, m this case, both J- 1 and J are the modes of the
dlstrlbutlon {c#J,},=
The lmphcation of the above results on the preferred bid and ask prices
charged by the market-maker IS lmmedlate P,, 1 P,(r*) > P,(r*) >PbJ, and
the preferred bid-ask spread IS always greater than the correspondmg
spread set by Garmans monopohst This also implies that the preferred bid-
ask prices straddle the market-clearmg price F at the mtersectlon of the
demand and supply curves P IS the unique price at which the expected rate
of buy orders equals the expected rate of sell orders (p = A), thus clearing the
market on the average Garman (1976, p 266) suggested that P IS the price
set by a benevolent zero-cost market-maker who provides a non-profit
public service by filling all mcommg orders from his inventory The choice of
P thus guarantees no expected drift m the market-makers inventory,
together with zero expected profit [under assumption (G)] However, if there
IS a fixed cost 5 per transaction, then the spread net of transaction cost ~111
46 Y Amthud and H Mendelson, Market-ma&q with muentory

be zero 5 Now, consider a market where there IS a competltlon among

market-makers The ordinary sufficient condltlons for perfect competition in
such a market are that entry 1s costless and free, that [see Fama (1970, p
387)] all available mformatlon 1s costlessly available to all market-makers,
and that they all agree on the lmphcatlon of current mformatlon for the
current demand and supply functions There may be additional costs to the
market-maker, notably the opportunity cost of his resources tied up m the
dealership activity, and the actual cost of transacting As usual, competition
leads to a pricing of the dealers services so as to reflect the costs of
provldmg these services If these costs were zero, competltlon would lead to a
zero spread, and the assumed homogeneity of expectations would lead to P,
=P, =P (since any other price will leave the market uncleared on the
average) The existence of positive costs of provldmg dealership services leads
to a positive spread which straddles P 6 When the cost per transaction 1s a
positive constant r [see dlscusslon followmg assumption (G)], the
competltlve spread will equal 5 > 0

4. The case of linear demand and supply

In this section we apply our general results to the special case of linear
demand and supply functions, D(P,) = y -6P,, and S(P,) = c(+/lP, Then,
R(P)=(~/@(YP--P*), and C(~)=(l/B)(~-Ia)
Here, the locus (12) of possible (&,p& IS the ellipse

+ A2IB
=g 9 (18)
whose axes comclde with the coordinate axes The likely rates are given by
the mtersectlon of the 45~line with the ellipse

1=p=JgB6/0 (19)

The rates chosen by Garmans monopolist, (Lm,pm),satisfy


whereas the market-clearing rates, (P,@), at which the demand and supply

5A posltlve spread ~111 exist whenever real resources are tied up m the market-makmg
6For a dIscussIon on the determmatlon of the bid-ask spread on the basis of the dealers cost,
see Demsetz (1968), West and Tmlc (1971)
Y Amthud and H Mend&on, Market-makmg with mventory 47

functions intersect and P, = P,, are


It follows that Iz~1 ~1 (and p <pm <@) (see also Theorem 3 8) The
correspondmg prices (usmg the respective upper-scripts) satisfy

PL<PrcP and Pf,>P:>F

It follows that the market-clearmg price IS contained m the interval [Pk,p,]

It IS of interest to investigate whether the bid and ask prices set bv the
market-maker always straddle the market-clearmg price P This happens If
and only if all I,, pk are not greater than A=pc (otherwise, d some I, >L
then P,, >P,, >Ir, if some /_+>@, then both prices are below P) This 1s
equivalent to requlrmg that both A,,, pMsIze =@ (since I, =max,..,,, 1S .MAk
and hzmaXk=O, 1,2, ,Mpk)
In our case, a suffclent condltlon for the market-clearing price P to be
straddled by all the bid-ask prices Pbk,P.& (k=O, 1, ,M) 1s


The suffclency of this condltlon follows since by Theorem 3 4 and the

defirutlons of R( ) and C( ),

g<R(r*)-C(r*)=(r*)2/6+(r*)2//l, (23)

130=@<P=2r* and pM=&$<@=2r*

Thus, when the (absolute values of the) slopes of the demand and supply
curves are not grossly different, P c [P,,, Pak] for all k
We now proceed to mvestlgate the behavior of the bid-ask spread set by
the market-maker According to the dynamic pnce/mventory adJustment
theory suggested by Smldt (1971, 1979), Barnea (1974) and Barnea and
Logue (1975), the spread should be mmlmal when the market-maker IS at his
preferred inventory level, and widens as his long or short posltlon 1s
undesirably high Our model yields a similar behavioral pattern
This phenomenon was attrrbuted to the higher risk of posltlomng See also Latane et al
(1975, pp 73-75), who m addltlon consldered the effect of rtsk on the spread set by dlfferent
speclahsts Benston and Hagerman (1974) and Stoll (1978b) provided a cross-sectlonal emplrxal
study on the effect of the risks Incurred by holdmg Inventory on the bid-ask swead across
48 Y Amrhud and H Mendelson, Market-makmg wzth mventory

The brd-ask spread correspondmg to supply and demand rates (a,~) IS

given by

Along the ellipse, we have

dd/dCc= (l/M/~ - I)

The function d(A, cc) 1s thus mnnmrzed at (A,#), and increases as (1,~)
approaches the hmrts (&,,O) and (0,~~) It follows (since pk increases with k)
that the market-maker reduces the bid-ask spread as he approaches the
likely inventory posrtron This result and Theorem 3 2 yield a bid-ask price
pattern which 1s illustrated m fig 3 Observe that this pattern resembles the
one presented by Latani et al (1975, p 74, fig 41)
Next, we study the effect of the market-makers inventory posrtron on the
total volume of transactrons By setting bid and ask prices he determines the
supply and demand rates, 1 and ~1,whose sum grves the expected number of
transactions per unit time Thus, Iz+/J represents the expected volume per
unit time [Equrvalently, (2 +p)- 1s the expected inter-transaction time ]
Using (24) we obtain

(d/d/M + P) = I - /W P/A

[~[__]___I___&__ .__ PO


J Inventory
(Profrrrrd poritaon 1

Fig 3 The bid-ask and the correspondmg spread, A, as a function of the market-makers
Inventory level
Y Amthud and H Mendelson, Market-makmg wtth tnventory 49

Thus, the volume 1s maxlmlzed at

As the inventory level approaches one of the short or long hmlts, the volume
decreases as expected
We conclude the dIscussIon of the hnear model by exammmg the
important issue of market efflclency In our dealership market, efficiency (m
the fan game sense) implies that it 1s lmposslble to make economic profits
by trading against the market-maker It 1s well known that a sufficient
condltlon for market efficiency 1s that market prices behave as a random
walk However, the prices m our model do not adhere to this property since
their dlstrlbutlon 1s mean-reverting * Thus, market agents might be tempted
to form a trading rule based on past price behavior, by which they would
profit through buying from the market-maker at low prices and selling back
to him at higher prices However, we shall show that any trading rule which
1s based on momtormg the behavior of the market-maker 1s useless and is
certain to produce a loss More specltically, we shall show that the pricing
pohcy of the market-maker results m all ask prices being greater than all bid
prices To show this, observe that by (23)

where the last mequahty follows from

The model also gives us the transient price behavior which derives from the transient
behavror of the inventory (recalhng that quoted pruzes are one-to-one related to Inventory)
Startmg from an uutlal state I at time t=O, the probablhty of lindmg the system in state 1 at
time th0 IS given by the (z,j)-entry of the matrix eA where A IS the correspondmg transItion-
rate matrix [see, e g , Howard (1971, ch 12)] The matrix e can be wrltten as the sum of a
matrix whose rows are all ldentlcal to the stationary probablhty vector (&,, c#J~,&, ,&) (row
ldentlty reflects independence of the mltlal state), and M addItIona matrices which reflect
dependence on the mltlal state The latter matrices are multlpbed by coefficients which are
exponentially decaying as a function of t Thus lf we observe the system at state I at some time,
the dependence of the state observed t time-units later on the mltlal state I dlmuushes to zero at
an exponential rate (as t-co) The corresponding relatlonshlp between observed quotations of
bid and ask prices and any mltml quoted prices readily follows Slmdarly, the observed
transaction prices form a tM-state birth and death process with hmitmg probabihtles
P{observed prlce=P,} =& P&+J.~), and P{observed prlce=P,} =I$& AJOlt+&) The
previous remarks regardmg temporal dependence apply as well to this process (with a dllferent
transition-rate matnx)
50 Y Amthud and H Mendelson, n4arket-makmg with inventory

Now, smce A0= @ and pLM

= ,/$, we have

Usmg expressron (20) for r* yields


which Imphes, for all k,J,

It has long been noted that the random-walk property of prices IS not a
necessary condltlon for market efficiency m the fair game sense [Fama
(1970)] In fact, emplrlcal studies have shown that serial dependence m price
changes co-exists with market efflclency m the sense that it IS lmposslble to
make profit by use of publicly available mformatlon Our model provides a
theoretical framework which implies a transaction-to-transaction price
dependence, together with market efficiency As can be seen from fig 3, the
systematlc pattern of prices cannot be used to make a profit since all ask
prices he above all bid prices Therefore, any trading rule that attempts to
profit from this price dependence 1s certain to produce a loss the bid-ask
spread will wipe out any prospective profit

5. Concluding remarks and possible extensions

The last result implies that market traders can make no profitable use of
mformatlon which 1s also available to the market-maker Even a knowledge
of the market-makers current inventory position and his pricing pohcy
(derived from the demand and supply functions) cannot produce a profitable
trading rule This result agrees with Bagehots (1971, p 13) observation that
the market-maker always gains m his transactlons with liquidity-motivated
transactors Yet, there may be insiders, 1 e , transactors who possess special
mformatlon which IS not available to the market-maker, and can make a
profitable use of it In other words, mslders have a more accurate assessment
of the demand and supply functions than that of the market-maker, and they
may use their superior mformatlon to make profit in excess of their cost
lmphed m the bid-ask spread As Bagehot (1971, p 13) noted, the market-
maker always loses to these insiders, and these losses represent an inventory
Y Amlhud and H Mendelson, Market-makmg wrth rnuentory 51

holdmg cost to the market-maker Clearly, our model IS structured to treat

the hquldlty-motivated transactors and not the insiders demand and supply
A model which ~111 take account of the mslders tradmg m an exphclt
manner is a most important extension This model should also contain a
learning mechamsm by which the market-maker uses market mformatlon to
update his assessment of the demand and supply [see Bagehot (1971, p 14)]
It IS worth relating our ObJectWe function (I e , expected average profit per
unit time) to value maximization, which implies here contmuous
dlscountmg of cash flows at some instantaneous discount rate u This gives
rise to a rate-dependent transaction-to-transaction discount factor, which
represents the present value of obtammg one-dollar at the next transaction
It has long been known [see Jewel (1963, pp 95&957)] that for small
discount rates (which are equivalent to close-to-umty transactlon-to-
transaction discount factors), the dlscounted value crlterlon turns out to be
well approximated by our average profit criterion I1 It should also be noted
that the relevance of dlscountmg to existing dealershlp markets 1s limited
when settlements take place a few days after the transactions, smce then the
actual tlmmg of a transaction 1s of no importance If the underlymg
condltlons m the market are such that dlscountmg 1s of importance, It 1s a
stralghtforward matter to formulate the problem as a discounted dynamlc-
programming problem Then, there 1s no direct analogy to our formulation
of the ObJectWe function (l), and our closed-form results will be replaced by
recursive relations Clearly, such a reformulation ~111be at the expense of the
models tractability
Consldermg the role of mventorles m this paper, note that our market-
makers pohcy depends on his stock inventory positIon, whereas his cash
flows appear only m the objective function The role of the cash position
may be interesting m a combined cash-mventory dependent pohcy, where the
market-maker maxlmlzes his expected utlhty of consumption
It may also be of interest to study the sensltlvlty of the results of our
model to the underlymg assumptions on the order arrival process This can
be done m several ways The order size may be assumed to be a random
variable which represents orders of varying size Furthermore. the
assumption of Poisson arrival may be extended to a more general renewal
process This may raise the necessity for an empirical study of the mterarrlval
time dlstrlbutlon *

See also Stoll (1978a, p 1144)

loThe need for contmuous dlscountmg results from the fact that the mtertransactton times are
not ldentlcally dlstrlbuted
If, for example, the yearly contmuous discount rate IS 14% (which IS equivalent to 15 Y0 per
annum) and the expected Inter-transactlon time 1s as large as an hour, the relevant discount
factor 1s 0 999984
A step m this dIrectIon IS Garbade and Lleber (1977)
52 Y Amlhud and H Met&son, Mark.&makmg with mwntory


Bagehot, Walter, 1971, The only game m town, Fmancnd Analysts Journal, March-Apnl, 12-14
Bamea, Anur, 1974. Performance evaluation of New York stock exchange specmhsts. Journal of
Fmanclal and Quantitative Analysis 9. 511-535
Barnea, Anur and Denms E Logue, 1975, The effect of risk on the market-makers spread,
Rnancml Analysts Journal 31, Nov /Dec. 45-49
Baumol, Wdham J , 1965, The stock market and economic etliclency (Fordham Umverslty Press,
New York)
Ek~a, Avraham and M Barry Goldman, 1977, On the dynanuc behavior of prices m
chsequhbrmm, Mlmeo
BcJ~, Avraham and M Barry Goldman, 1978, Market pricts vs eqmhbnum prices Returns
variance, serial correlation, and the role of the specmhst, Working Paper no 78-51 (New
York University, New York)
&~a, Avraham and Nds H Hakansson, 1977. Some alternative market mechanisms, Workmg
Paper no 104 (New York Umverstty, New York)
Bela, Avraham and Nils H Hakansson, 1979, On the feasiblhty of usmg a mechanical speciahst
m securities trading, Paper presented at the XXIV Meeting of The Institute of Management
Benston, George J and Robert L Hagerman, 1974, Detemunants of bid-&ked spreads m the
over-the-counter market, Journal of Financial Economics 1, 353-364
Cohen, Kalman J, Walter L Ness, Hltoshl Okuda, Robert A Schwartz and David K
WhItcomb, 1976, The determinants of common stock returns volatdrty An mternatlonal
companson, Journal of Fmance 31, 733740
Cohen, Kalman J, Steven F Maler, Walter L Ness, Hitoslu Okuda, Robert A Schwartz and
David K Whitcomb, 1977, The impact of designated market makers on security prices I,
Empwlcal evidence, Journal of Banking and Finance 1, 219-235
Cohen, Kalman J, Steven F Mater, Robert A Schwartz and David K Whitcomb, 1977a, The
Impact of designated market makers on security prices II, Pohcy proposals, Journal of
Banking and Finance 1,236-247
Cohen, Kalman J, Steven F Marer, Robert A Schwartz and David K Whitcomb, 1977b, On
the existence. of serial correlation m an e&lent securltles market Working Paper no 109
(New York University, New York)
Cohen, Kalman J, Steven F Maler, Robert A Schwartz and David K Whitcomb, 1978, The
returns generation process, returns variance and the effect of thinness m securities markets,
Journal of Finance 33, 149-167
Demsetz, Harold, 1968, The cost of transactmg, Quarterly Journal of Economics 82, 3353
Fama, Eugene F, 1970, Eficient capital markets A review of theory and empirical work,
Journal of Finance 25, 383-417
Garbade, Kenneth and ZVI Lleber, 1977, On the independence of transactions m the New York
Stock Exchange, Journal of Banking and Finance 1, 151-172
Garman, Mark B , 1976, Market microstructure, Journal of Fmancial Economics 3, 257-275
Howard, Ronald A, 1971, Dynamic probablhstic systems, Vol II Semi-Markov and declslon
processes (Wiley, New York)
Jewell, Wdham S, 1963, Markov-renewal programmm& II Infinite return models, Example,
Operations Research 1I, 949-971
Latane, Henry A, Donald A Tuttle and Charles P Jones, 1975, Security analysis and portfolio
management (Ronald Press, New York)
Logue, Denms E, 1975, Market-making and the assessment of market efliciency, Journal of
Fmance 30 115123
Smldt, Seymour, 1971, Which road to an efficient stock market Free competltlon or regulated
monopoly, Fmancial Analysts Journal 27, Sept /Ott , 18-20, 64-69
Smldt, Seymour, 1979, Contmuous versus mtermlttent trading on auction markets, Paper
presented at the Western Fmance Assoctatlon
Stall, Hans R , 1976, Dealer Inventory behavior An empirIcal mvestlgatlon of NASDAQ stocks,
Journal of Fmancial and Quantltatlve Analysis, 356380
Y Amrhud and H Mendclson, Market-makmg wtth mwntory 53

Stall, Hans R , 1978a, The supply of dealer servtccs m securltlcs markets, Journal of Fmance 33,
Stall, Hans R, 1978b, The pncmg of security dealer servxes An empIrIcal study of the
NASDAQ stocks, Journal of Fmance 33, 1153-l 172
Tnuc, Seha M, 1972, The economics of bquldlty servxcs, Quarterly Journal of Economics 86,
West, Richard R and Seha M Tmlc, 1971. The economxa of the stock market (Praeger, New